An open benchmark published in Cell, LongevityBench, evaluated 18 frontier AI systems and five compact models trained specifically on aging data across 17 tasks spanning clinical, genetic, epigenetic, transcriptomic and proteomic measurements. The specialist models, ranging from 0.6 to 9 billion parameters, matched or exceeded far larger general-purpose systems when reasoning directly from primary biological data. Fluency in the aging literature proved a poor proxy for the ability to extract biological meaning from unfamiliar omics profiles.
Key Points
- Compact aging-trained models matched or beat far larger frontier systems
- Benchmark uses 25,457 prompts built from primary biological measurements
- Biological age prediction remains difficult; 328 proposed targets need validation
Longevity Analysis
Interpretation, not information, is the bottleneck in aging biology. Methylation patterns, plasma proteomes and transcriptomic shifts are the language through which tissues report on their own repair capacity, inflammatory load and metabolic state, and reading that language requires models grounded in the biology rather than in commentary about it. Open release of the benchmark, models and platform allows the resulting hypotheses to be tested against experimental biology instead of accepted on the strength of fluent output — the same discipline that separates a plausible intervention from a validated one at the individual level.
Original published by Longevity.Technology, by Eleanor Garth.